Antibody-Drug Conjugates (ADC) Registration Document

Data Characteristics

Antibody-Drug Conjugate (ADC) registration documents involve diverse and complex data sources. These include preclinical study reports (pharmacology, toxicology, pharmacokinetics), CMC (Chemistry, Manufacturing, and Control) files, clinical trial protocols, investigator brochures, clinical study reports, statistical analysis plans, and post-market risk management plans. Data updates are frequent, especially during clinical trials, where data is continuously generated and revised as trials progress. Document structures typically follow ICH guidelines, such as the CTD format, organized into modules. Fields cover molecular structure, purity, stability, in vitro and in vivo efficacy, safety indicators (e.g., AUC, Cmax, LD50), clinical efficacy endpoints (ORR, PFS, OS), and units such as mg/kg, µg/mL, nM, days, months, years. These documents often contain complex biological and statistical terminology.

Constraints on Deployment and Upgrade

The complex data characteristics of ADC registration documents impose specific requirements on FastGPT's deployment and upgrade processes. First, the large volume of data and frequent updates demand high throughput and scalability from the underlying storage, along with support for rapid index updates to handle continuous clinical data entry and version iterations. Second, diverse document formats (PDF, Word, Excel, images) and nested structures require robust file parsing capabilities and intelligent segmentation strategies to ensure accurate and complete information extraction. This is particularly critical for files containing numerous charts and biological sequence information, to avoid parsing errors or the omission of key data. Furthermore, the recognition of specialized terminology and units tests the model's understanding capabilities and the granularity of knowledge base construction, impacting recall effectiveness. During upgrades, ensuring data migration integrity and model compatibility is essential to prevent knowledge base invalidation or degraded query performance due to version differences.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`UPLOAD_FILE_MAX_SIZE`	`1000 MB`	Individual ADC registration files (e.g., clinical study reports) are often large, requiring support for large file uploads.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Parsing complex PDFs or scanned documents can be time-consuming; sufficient time prevents timeout interruptions.
`Chunk size` (Segment Length)	`800–1200 characters`	Balances semantic completeness and retrieval efficiency, avoiding excessive fragmentation or information overload.
`Similarity threshold` (Similarity Threshold)	`0.78`	ADC data is highly specialized; a higher threshold reduces recall of irrelevant content, improving accuracy.
`Recall count` (Recall Count)	`Top 5`	Prioritizes the most relevant results, reducing the engineer's filtering workload.
`maxContext`	`8192`	Ensures large sections of clinical data or CMC descriptions are fully included in the context, improving understanding.

Common Mistakes

Inaccurate knowledge base query results after an upgrade. This occurs because the new model version changes how certain specialized terms are embedded, leading to vector matching deviations.
Long periods of unresponsiveness or HTTP 504 Gateway Timeout errors when uploading large PDF files. This often happens when the file parser's processing time exceeds the default gateway timeout limit.
After deployment, tabular data in some clinical trial reports is not extracted correctly, appearing as empty fields or misaligned data. This is due to the file parser's insufficient capability to handle complex table structures.

Verification

Upload several typical ADC registration documents (e.g., pharmacokinetic reports, clinical trial protocols). Check file parsing progress and knowledge base construction status. Confirm no errors and that processing time is within a reasonable range.
Query key data points in the knowledge base (e.g., a drug's Cmax value, specific adverse event rates) multiple times. Compare the answers with the original text to confirm recall accuracy.
Simulate typical engineer query scenarios by asking questions containing specialized terminology and abbreviations. Observe the recalled results and cited original text to verify if the Similarity threshold (Similarity Threshold) and Recall count (Recall Count) effectively locate the required information.

Note: The values provided are common starting points. Measure them against your own samples.

Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.